Ask a model for fifty headlines and you will get fifty headlines. They will arrive in seconds, spelled correctly, on brief, each one plausible. This is the moment most people decide the creative problem is solved. It is also the moment the actual problem starts: which one ships?
Fifty plausible options is not an output. It is a decision waiting to happen. And the decision is the part the model did not do.
The demo and the campaign
A demo needs to be impressive once. A campaign needs to be right, repeatedly, in public, with money behind it. The distance between those two is not capability. It is judgment. Which claim leads. Which word is doing work and which is decoration. Whether the joke lands with a cold audience or only with the people who wrote it. Whether this concept can survive fourteen variations without going stale.
Models have made competence free. Anyone can now produce work that looks professional. Which means looking professional is worth nothing, and the entire remaining game is selection.
What taste actually is
Taste gets talked about like a gift. Operationally it is much less romantic. Taste is the ability to kill work quickly and be right about it.
It is a set of standards you can articulate. Why this headline and not that one, in words, not vibes. It is pattern recognition earned from seeing thousands of pieces of work meet thousands of audiences and remembering what happened. It is knowing that the client room applauds one thing and the market rewards another, and never confusing the two again.
I spent years as a Creative Director defending work in rooms. What that trains, more than anything, is the ability to explain a judgment. Not "I like it" but "this works because the first three words carry the claim and the image does the proof." The explanation matters, because a judgment you can explain is a judgment you can teach. And a judgment you can teach, it turns out, is a judgment you can encode.
Taste as software
This is the part I did not expect. Taste can be written down, and once it is written down, machines can apply it.
The review bench I run does not generate copy. It judges it. Multiple agents, each holding a different piece of the standard: one carries the banned claims and the compliance lines, one carries the voice, one carries what strong work looks like against what weak work looks like, with examples of both. Copy goes in, argument comes out, and most of it dies before a human ever reads it. The human taste gets spent where it matters, on the short list, not on the flood.
The same applies to swipe intelligence. A library of work that already won, organized by why it won, is taste in storable form. A model with access to it makes decisions a model without it cannot.
None of this removes the human. It concentrates them. My standards run all day without me; I spend my attention updating the standards.
The compounding asset
Here is the asymmetry worth building a career on. Capability commoditizes. Every quarter, the thing that was hard for a model becomes easy, and everyone gets it at the same time, at the same price. Nothing you rent from a vendor is an advantage for long.
Judgment compounds. Every campaign that runs teaches you something the next one uses. Every standard you write down makes the system slightly harder to copy, because the standards came from your losses, and nobody else has your losses.
So when people ask what to hire for, or what to build first, the answer is the same. The generation is handled. Invest in the thing that decides. Taste is the bottleneck, and the bottleneck is always where the value pools.